This paper aims to test an important hypothesis in financial economics: whether equity returns are predictable over various horizons? The conventional wisdom in the literature is that aggregate dividend yields strongly predict excess returns, and the predictability is stronger at longer horizons (Fama and French (1988), Campbell (1991), and Cochrane (1992)). In contrast, Ang and Bekaert (2007) find that dividend yields, together with the short rate, predict excess returns only at short horizons, and do not have any long-horizon predictive power. In this paper, we undertake an analysis of both in-sample and out-of-sample tests of equity return predictability to better understand the empirical evidence on return predictability over different time horizons. We first propose a nonparametric test to examine the predictability of equity returns, which can be interpreted as a signal-to-noise ratio test. Our empirical results show that the short rate, dividend yields and earnings yields have good predictability power for both short and long horizons, which is different from both the conventional wisdom and Ang and Bekaert (2007). Also, using our nonparametric test, a comprehensive in-sample and out-of-sample analysis documents that the predictor variables (dividend yields, earnings yields, dividend payout ratio, short rate, inflation, book-to-market ratio, investment to capital ratio, corporate issuing activity, and consumption, wealth, and income ratio) have predictability power on equity returns but this cannot be well captured by linear prediction models. In addition, we use the nonparametric test to compare the conventional long-horizon prediction regression models on predictor variables with the historical mean model, where there has exists a debate about which model has better forecasting power for equity returns (Campbell and Thompson (2007) and Goyal and Welch (2007)). We find that the prevailing prediction model has a better forecasting power than the historical mean model because the former has a lower neglected signal-to-noise ratio. Finally, our nonparametric predictive models have lower RMSE than the historical mean model at both short-horizon and long-horizon. Using our nonparametric methods, both combined and individual forecast outperform the historical average.